Industrial Time Series Prediction Considering Data Uncertainty

A large number of key variables are contained in the industrial production process,and predict precisely of variables in the production process can provide guidance information for decision-making of industrial systems.The large amount of real-time data collected by SCADA system provides support for the establishment of data-based model.However,missing points and high noise are generally in the collected industrial data due to sensor failures,storage anomalies and equipment problems,which poses a huge challenge in ensuring the accuracy of modeling.In view of the two uncertainty factors of missing points and high noise in industrial time series species,the industrial time series variable prediction method considering data uncertainty is studied in this paper.The specific contents are as follows:For the data imputation problem of incomplete data caused by missing points in the data,a multiple imputation algorithm based on Bayesian principal component analysis and bootstrap is proposed for data imputation in time series.The proposed method is composed of three steps.Firstly,we need to sample n times with replacement through the bootstrap method to prepare for the subsequent multiple imputation.Then,the Bayesian method combined with principal component analysis and the Expectation Maximization algorithm is utilized to estimate the model parameters and missing points for each data subset.Finally,the estimated value can be obtained by comprehensive analysis of these groups of filling values.To test the validity of the proposed method,three kinds of interpolate methods are chosen to make comparative experiments,synthetic data sets,benchmark data sets and real gas data sets of steel enterprises as experimental data are employed in this paper.Experimental results show that the imputation precision of the raised method is better than that of the counterparts.For the modeling problem of input uncertainty caused by high noise in the data,an Echo State Network prediction model considering input noise is proposed for time series prediction,where the state of the dynamical reservoir in the ESN model is linearized locally by extended Kalman filter.In the phase of model parameters determination,the expectation maximization algorithm is utilized to update all the uncertain parameters iteratively to construct the prediction interval,where the forward-backward algorithm is adopted for the state estimation.To verify the effectiveness of the raised method,the standard ESN method and classical time series prediction methods considering input noise are used for comparative experiments,synthetic data sets and real industrial gas data sets of steel enterprises are tested in this paper.Experimental results show that the prediction accuracy of the raised method is better than that of the counterparts.